Feb 26, 2026

Daily

Google API keys weren't secrets, but then Gemini changed the rules

• Google API keys, once deemed safe for public use, now pose a significant risk as they can access sensitive data through the Gemini API, with nearly 3,000 keys identified as vulnerable across the internet.

Nano Banana 2: Google's latest AI image generation model

• Nano Banana 2 merges the advanced capabilities of Nano Banana Pro with the speed of Gemini Flash, enabling rapid image generation and editing across Google products.

Making MCP cheaper via CLI

• MCP's method of loading tool schemas incurs excessive token costs, while CLI's approach reduces this by 94% through on-demand discovery, significantly lowering operational expenses.

Large-Scale Online Deanonymization with LLMs

• LLMs can deanonymize users with high precision by analyzing a few online comments, revealing personal details like location and interests, and effectively searching for identities across platforms such as Hacker News and Reddit.

The First Fully General Computer Action Model

• FDM-1 is the first fully general computer action model, capable of performing complex tasks like CAD modeling and real-world driving at 30 FPS, trained on an unprecedented 11-million-hour video dataset.

Learnings from 4 months of Image-Video VAE experiments

• Key finding: Linum's experiments reveal that better compression in Variational Autoencoders (VAEs) does not guarantee improved downstream generation quality, highlighting the complexity of balancing reconstruction fidelity with generative performance.

SynthID

• SynthID embeds an invisible digital watermark in AI-generated images and videos, ensuring content integrity without compromising quality during creation.

Show HN: ZSE – Open-source LLM inference engine with 3.9s cold starts

• ZSE is an ultra memory-efficient LLM inference engine that leverages the Intelligence Orchestrator to optimize performance based on available memory, enabling the execution of large models with minimal resource usage.

• LLM agents can accurately deanonymize users from anonymous online posts, identifying individuals with high precision across platforms like Hacker News, Reddit, and LinkedIn, even with minimal data.

America, and probably the world, stands on a precipice

• America faces a critical juncture as Secretary Hegseth demands unrestricted military access to Anthropic’s AI, potentially enabling autonomous weapons and mass surveillance without human oversight.

PA bench: Evaluating web agents on real world personal assistant workflows

• PA Bench is a new benchmark that evaluates web agents on realistic, multi-step personal assistant workflows, addressing the limitations of existing benchmarks that focus on isolated tasks.

Steering interpretable language models with concept algebra

• Steerling-8B allows for concept algebra, enabling users to manipulate human-understandable concepts during inference, thus enhancing model output control without the need for retraining or prompt engineering.

Ferret-UI Lite: Lessons from Building Small On-Device GUI Agents

• Ferret-UI Lite is a compact, end-to-end GUI agent that excels across platforms, achieving competitive performance through innovative techniques like chain-of-thought reasoning and reinforcement learning.

• Universal vulnerability in open-weight LLMs was demonstrated, with attack success rates nearing 100% across 50 models using 23 distinct strategies, revealing a critical flaw in model safety mechanisms.

Lidar waveforms are worth 40x128x33 words

• Lidar's potential is enhanced through a learned DSP that processes full waveforms with a transformer architecture, improving point cloud fidelity by leveraging adjacent waveform features.